What question did this study set out to answer?

March 17, 2026Chemical Engineering Journal Advances0 citationsOpen Access

Accurate Prediction of Ideal Gas Heat Capacity Cp(T) Using Machine Learning: Data Augmentation and Ensemble Learning for Robust Thermodynamic Modeling

RBRoda BounaceurCentre National de la Recherche Scientifique FPFrancisco Carlos PaesCentre National de la Recherche Scientifique JJJean-Noël JaubertCentre National de la Recherche Scientifique

Key Points

To develop a robust model for predicting temperature-dependent ideal-gas molar heat capacities using machine learning approaches.
Used Mordred descriptors to compute molecular features.
Applied artificial neural networks (ANNs) for prediction.
Augmented dataset with correlations for 1,471 compounds.
Trained multiple ANN models using bagging ensemble configuration.
Achieved high predictive accuracy (R² > 0.999) for both training and testing datasets.
Demonstrated strong generalization across various chemical spaces.
Provided continuous predictions for ideal-gas heat capacities across relevant temperature ranges.

Abstract

• Flexible framework using Mordred descriptors and neural networks for Cp prediction. • Cp normalized by infinite-temperature limit improves dataset balance. • High accuracy achieved (R² > 0.999) with bagged neural network ensemble. • Model generalizes well across chemical space, including complex molecules. Reliable predictions of ideal-gas molar heat capacities at constant pressure are a cornerstone of thermodynamic modeling. In the absence of a reliable model or correlation for the estimation of the ideal-gas heat capacity, an equation of state (EoS) alone cannot ensure thermodynamically consistent calculations of state-function variations, nor can it provide real-fluid heat capacities, since it only yields residual contributions. Changes in internal energy, enthalpy, entropy, exergy, and real-fluid heat capacities require the integration of the ideal-gas contribution which depends solely on the temperature-dependent ideal-gas heat capacity. Thus, coupling an EoS with an ideal-gas heat capacity model enables full thermodynamic property evaluation in process simulators. In this paper, we propose a deep-learning framework for continuous prediction of temperature-dependent ideal-gas molar heat capacities at constant pressure ( C p ) via artificial neural networks (ANNs) trained using molecular descriptors computed with Mordred and a high-quality experimental database comprising 1,471 compounds. The dataset was augmented using experimentally validated correlations, generating 100 C p data points per molecule over their corresponding temperature range and several ANN models were trained in a bagging ensemble configuration. The resulting model achieves excellent predictive accuracy (R² > 0.999 for both training and test sets) while demonstrating strong generalization to external datasets, indicating that the framework captures transferable structure-thermodynamic relationships. By explicitly treating temperature as an input variable, the proposed model provides smooth and continuous ideal-gas C p ( T ) predictions while preserving physically consistent thermodynamic trends across temperature ranges relevant to industrial applications. The developed prediction tool is available in a web interface (link).

问 AI

Bookmark

View Full Paper